Anda di halaman 1dari 11

ARTICLE IN PRESS

Physica A 380 (2007) 418428


www.elsevier.com/locate/physa

Music walk, fractal geometry in music


Zhi-Yuan Sua, Tzuyin Wub,
a

Department of Information Management, Chia Nan University of Pharmacy & Science, Tainan 717, Taiwan, ROC
b
Department of Mechanical Engineering, National Taiwan University, Taipei 106, Taiwan, ROC
Received 16 November 2006; received in revised form 5 January 2007
Available online 1 March 2007

Abstract
In this study, sequences of musical notes from various pieces of music are converted into one-variable random walks
(here termed music walks). Quantitative measurements of the properties of each musical composition are then performed
by applying Hurst exponent and Fourier spectral analyses on these music-walk sequences. Our results show that music
shares the similar fractal properties of a fractional Brownian motion (fBm). That is, music displays an anti-persistent trend
in its tone changes (melody) over decades of musical notes; and music sequence exhibits generally the 1=f b -type spectrum
(fractal property), with apparently two different b values in two different temporal scales.
r 2007 Elsevier B.V. All rights reserved.
Keywords: Music walk; Fractal; Fractional Brownian motion (fBm); Long-range correlation; Hurst exponent; Power spectrum

1. Introduction
There are almost as many styles of music as there are composers, as each composer imparts their own
creative preferences and ideas when working on a composition. However, beyond this variety and
individuality, are there rules or an underlying structure that essentially differentiates a musical work from
meaningless collections of notes? This hypothesis can be investigated through quantitative analysis of the
fractal properties of musical compositions. Although the mathematical theory of music is rather deep and
complicated [1], quantitative study can nevertheless be performed by appealing to the notion of fractional
Brownian motion (fBm) and Fourier spectral analysis. Interestingly, as disclosed in this study, the results show
that music exhibits the ubiquitous property of long-term correlation (fractal) over decades of notes in two
apparently different scaling ranges. This underlying structure may explain why music sounds generally
pleasing, and shows how music imitates the harmony of nature.
2. Literature review
Music is conventionally dened as an ordered arrangement of sounds of different acoustic frequencies
(pitches, tones) in succession (melody), of sounds in combination (harmony), and of sounds spaced in
Corresponding author. Tel.: +886 2 33662708; fax: +886 2 23631755.

E-mail address: tywu@ntu.edu.tw (T. Wu).


0378-4371/$ - see front matter r 2007 Elsevier B.V. All rights reserved.
doi:10.1016/j.physa.2007.02.079

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

419

temporal succession (rhythm) [2]. An individual pitch (or tone) does not by itself produce music; it is the
melody (successive changes in pitch) and rhythm (successive changes in tone duration) that are the two
essential elements of music. The arithmetical relationships of notes in harmony was speculated upon at length
by numerologists in the middle Ages, and it has been known for thousands of years that notes of a frequency
ratio 2:1 produce the octave. Much of the theoretical and experimental work about music and tuning was
performed in between 15th and 17th centuries. For example, in dividing an octave, one of the most popular
methods considered desirable (though not essential) in order to play a musical composition harmonically in all
keys is the equal temperament rule [1,2]. If f 0 denotes the pitch (frequency) of the base note (e.g. middle C, a
note an octave higher than the base note has a pitch ratio f =f 0 2 with reference to the base note. In the equal
temperament rule, an octave is divided into 12 geometrically equal intervals in pitch (i.e., the 7 white and 5
black keys on a piano) according to:
f i =f 0 2i=12 .

(1)

Thus a semitone is represented by the pitch interval i 1, a full tone by i 2, a minor third by i 3, a major
third by i 4, a fourth by i 5, a diminished fth by i 6, a sonorous fth by i 7 and an octave by i 12.
In such a calibration, each note has a frequency ratio (with reference to the base note) of approximately a ratio
of small integers. For instance, a fourth (i 5) has a ratio of 25=12 1:3348  43; a fth (i 7) has a ratio of
27=12 1:4983  32. The exception is the diminished fth (i 6), which has a ratio 26=12 1:4142  1000=707,
which is obviously not a ratio of small integers. This note interval has traditionally been considered dissonant
and thus has rarely been used in classical compositions [2].
Recent analysis of musical structure has revealed evidence of a long-range scaling property similar to that
found in natural landscapes, such as the proles of mountains, coastlines, etc. [3]. By performing power
spectral analysis upon the audio signals recorded from various selected pieces of different music styles
(classical, rock, jazz and blues), Voss and Clarke [4,5] have shown that the power spectra of both the
instantaneous loudness and the frequency uctuations (roughly approximating to the melody) of music vary
approximately as 1=f , a typical characteristic of a scaling noise found in electronic components [6,7]. Such a
tendency has also been conrmed by Schroeder [8] and Campbell [9], whose results imply the existence of a
certain long-range correlation, uctuating in quantity according to piece and style, over a major part of the
music.
That music shows a 1=f -spectrum is not greatly surprising, because 1=f 0 -dependence signal (random noise)
sounds meaningless, while 1=f 2 -dependence signal (Brownian noise) would sound a little bit dull [10].
Therefore, 1=f -noise shows a greater correlation in adjacent signal values than random noise and a weaker
correlation when compared with Brownian noise [11]; and the property of 1=f -dependence signal appears to
serve as a compromise between too little and too much surprise, randomness or musicality in a musical
composition. Music is a blend of randomness and orderliness. Consequently, 1=f -signal is a good candidate
for stochastic music composition [1215].
Several researchers have made use of quantitative methods in order to study the properties of music. Hsu
and Hsu [16,17] (see also Ref. [2]) analyzed the variations in pitch interval (the i dened in Eq. (1)) between
successive notes in a series of music scores composed by Bach and Mozart, and were able to show that the
incidence frequency (the frequency of appearance of each pitch interval) F approximately shows a power-law
relationship, F / iD . The value of the exponent D for each different score varies between 1 and 3 and is not
an integer; in other words, the incidence of change in acoustic frequency in a section of a music score exhibits
fractal geometry. The authors also considered the use of this fractal property in an attempt to reduce the
length of a composition while still being able to maintain its style.
In a recent study, Shi [18] applied two different types of correlation analyses, frequency dependent and
frequency independent, to examine the music sequences converted from folk songs and piano pieces. His
results indicated that music sequences have a long-range power-law behavior in both analyses, and that the
fundamental principle of music is the obtaining of a balance between repetition and contrast. Further,
Bigerelle and Iost [19] applied the variance method to study the fractal dimensions in 180 musical works of
various styles. Based on their statistical results, they proposed that various music pieces could be categorized
by fractal dimension. Madison [20] used a similar approach to study different musical scores, and found that
the Hurst exponent plays an important role in examining the emotional expression of a musical performance.

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

420

The study by Manaris et al. [21] of a 220-piece corpus (including baroque, classical, romantic, 12-tone, jazz,
rock, DNA strings, and random music) revealed that esthetically pleasing music might be describable under
the ZipfMandelbrot law. Gunduz and Gunduz [22] studied the mathematical structures of six songs by
treating them as complex systems. The key approach of their analysis was to calculate the fractal dimension of
a scattering diagram constructed from the melody of the six songs.
In this paper, the sequence of musical notes is converted into a one-variable random walk (here termed the
music walk). The resulting graph appears to resemble a mountainous landscape in prole view, with jagged
ridges of all lengths (scales), from minor bumps to enormous peaks. The fractal property of the music walk is
then explored by adopting the notion of fBm proposed by Mandelbrot and Van Ness [23], whose description is
given in the following section.
3. Properties of fBm
One common method for extracting hidden structure information from a uctuating time-series in signal
processing is to calculate its power spectrum distribution. By applying the Fourier transform to a given
continuous time series xt, the time-distribution signal is converted into a frequency-distribution one
Z
xf / xtei2pft dt
(2)
and the power spectrum of the signal xt is obtained as
Sx f / jxf j2 .

(3)

Another useful measurement of a time series is the autocorrelation function, C x t, which is dened as
C x t hxt  xt ti,

(4)

where hi denotes averaging the quantity over time t. The autocorrelation function is a measure of how well the
data set correlates with itself after a time lag t. The power spectrum and autocorrelation function of a time
signal are not independent; they are related by the WienerKhintchine relations [24]
Z
Sx f / C x t cos2pf t dt,
(5)
Z
C x t /

S x f cos2pf t df .

(6)

The most commonly encountered uctuating signal is the normally distributed random noise (also termed
Gaussian white noise), wt. The autocorrelation function for such noise is a delta function dt; that is,
successive signals in random noise are totally uncorrelated. The power spectrum distribution of random noise
is at (Sw f / f 0 , a constant), indicating an equal composition of components of all frequencies. Integrating
random noise with time then results in a new uctuating time signal,
Z
xw t wt dt,
(7)
where xw t can be regarded as the instantaneous location of a walker who takes steps to the right or left
along a one-dimensional line, with each step size determined by the successive signal values of a random noise
wt. Such motion is termed a random walk or Brownian motion, named after Robert Brown, a Scottish
botanist. It is a well-known fact in statistics that the average distance traveled within a time period T for a
random walk is given by the diffusion law
Dxw T hjxw t T  xw tj2 i1=2 / T 1=2 .

(8)

The power spectrum of Brownian motion is therefore S xw f / 1=f 2 . In contrast to random noise wt,
successive values in a random walk signal xw t are strongly correlated.

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

421

Mandelbrot and Van Ness [23] generalized expression (8) into the form
DxH T hjxH t T  xH tj2 i1=2 / T H ,

(9)

with 0 oHo1, H being the Hurst exponent. The corresponding motion (time signal) xH t is now generally
termed fBm. For H41=2, the graph of fBm xH t is less rugged-looking, or smoother, than that of Brownian
motion itself (H 12); and xH t tends to increase (decrease) in the future if it is increasing (decreasing) in the
past, i.e., showing the property of persistence in the increments of signal values Dx. For Ho12, the graph of
xH t is more rugged-looking and less smooth than that of Brownian motion; and xH t tends to decrease
(increase) in the future if it is increasing (decreasing) in the past, i.e., showing a trend of anti-persistence in the
increments of signal values Dx. When H 12, the signal is restored to the conventional Brownian motion,
xw t, and increments of adjacent signal values (not the signal values themselves) in a Brownian motion are
totally uncorrelated (random). The power spectrum of an fBm has the property
SxH f / 1=f b

(10)

with the power b related to the Hurst exponent H by the relationship b 2H 1 [25,26], and 1obo3 for
0oHo1.
According to the WienerKhintchine relations, the autocorrelation function cxH t of the time series xH t
having a 1=f b power spectrum with b close to 1 also exhibits a power-law decaying behavior. The slowly
decaying autocorrelation function implies that the time series are self-correlated over a rather long period of
time; that is, the signal at any time is still related in a certain way to a signal appearing long before. The time
signal is scale invariant; such a uctuating signal is simply the time analogy of many naturally found selfsimilar geometries, which are now categorized under the name fractals. The ne structure of fractal
geometry, when magnied, looks like the structure as a whole, and in a fractal time signal, a short time section
of the signal has the same statistical properties as the whole signal. Theoretically, the fractal dimension (boxcounting dimension) D of the landscape traced by one-variable fBm and its Hurst exponent H is related by
D 2  H [2529].
4. Music walk as an fBm
Choosing an arbitrary note (e.g. c, a note an octave below middle C as a base note, all notes in a score can
then subsequently be digitized and labeled by the index i, as dened in the equal temperament rule, for
example, c with i 0, # c with i 1, d with i 2; . . . etc. Also, similar to the work of Shi [18], when
considering the difference in tone duration of each note, the shortest time in the score is chosen as the step unit
1
of the time-wise coordinate. That is, if the shortest time (beat) in a music score is 16
note, then the index value
1
representing the pitch of a 4 note is repeated 4 times when constructing the time sequence from that music
piece. The entire music score is thus converted into a reminiscence of a random walk (music walk), xn. The
sequence xn is simply the discrete version of the continuous time signal xt, with n being the number of notes
(in multiplicity of the note with smallest time, as explained above) representing the elapsed time in a musical
movement, and x being the value of the index i, which represents the pitch of the note. Therefore, both the
melody (changes in pitch) and the rhythm (changes in tone duration) of a composition are taken into account
in the digitized sequence xn. Note that Hsu and Hsu [16,17] did not include any effect of tone duration in
their analysis. The typical xn curves are shown in Fig. 1; these are the resulting curves from the conversion of
the violin scores from Gavotte written by Gossec, Le Cygne written by Saint-Saens and Ave Maria written by
Bach and Gounod. As can be seen from Fig. 1, these plots resemble the fractal proles of skyscrapers, or
mountain ridges if x and n are properly scaled.
Characteristic of the uctuating signal xn is then studied by calculating the Hurst exponent H and the
power b of the power spectrum distribution for various music-walk sequences converted from various music
pieces: (I) Gavotte by Gossec; (II) Le Cygne by Saint-Saens; (III) Ave Maria by Bach and Gounod; (IV)
Beethoven, Op. 24, No. 5, Sonata for piano (IVa) and violin (IVb); (V) Mozart, Op. 70, No. 1, duo violins,
violin 1 (Va) and violin 2 (Vb); (VI) Pachelbels Canon and Gigue, 3 violins and continuo, violin 1 (VIa) and
violin 2 (VIb); (VII) Bach, J. C., six duets, No. 1, duo violins, violin 1 (VIIa) and violin 2 (VIIb). In order to
avoid the difculty that might be encountered in selecting the proper key note from a chord, the music scores

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

422

200
175

Gavotte
Le Cygne
Ave Maria

150

x (n)

125
100
75
50
25
0
0

100

200

300
n

400

500

600

Fig. 1. Music-walk sequences obtained from the violin score of three different pieces of music, Gavotte, Le Cygne and Ave Maria. For
clarity, the data values of Le Cygne and Gavotte have been shifted uniformly upward by 40 and 80 units, respectively.

Table 1
Values of the Hurst exponent H and exponent b of the power spectrum obtained from various music-walk sequences (short time-interval
analysis)
Short time-interval
analysis

II

III

IV(a)

IV(b)

V(a)

V(b)

VI(a)

VI(b)

VII(a)

VII(b)

H
b
2H 1
Deviation

0.35
1.80
1.70
5.88%

0.38
1.67
1.76
5.11%

0.22
1.43
1.44
0.69%

0.38
1.74
1.76
1.14%

0.34
1.83
1.68
8.93%

0.22
1.50
1.44
4.17%

0.22
1.64
1.44
13.89%

0.23
1.57
1.46
7.53%

0.26
1.55
1.52
1.97%

0.32
1.54
1.64
6.10%

0.30
1.51
1.60
5.63%

we choose to study in this paper are all violin scores (except IVa). The Hurst exponent H is obtained by taking
logarithm of both sides of the discrete version of Eq. (9) to give
log sDx K H log jDnj,

(11)

where sDx is the standard deviation of the increments in pitch Dx corresponding to the interval size (number
of notes) Dn, and K is a constant. Similarly, the b value is just the slope of the straight line obtained from leastsquare curve tting of the power spectrum (in logarithmic scales) of the music-walk sequence xn. In this
study, power spectra are obtained by applying discrete version of the Fourier transform (2) to music-walk
sequences. Many commercialized numerical software packages provide subroutines for calculating power
spectrum from given time sequence, such as MATLAB.
The results are summarized in Tables 1 and 2. Graphical representations of the results from the rst three
music pieces I, II and III are shown in Figs. 24, respectively. Part (a) of each gure shows the root-meansquare deviation of the music sequence, sDx; and part (b) is the power spectrum of the sequence. The sharp
rise or drop in the values of sDx near the end of the sequence reects the nite-number-of-data effect and
should be excluded from analysis. The solid lines superimposed on the graphs are obtained from least-square
curve tting of the data points in the appropriate intervals, and whose slopes represent respectively the Hurst
exponent H of the graph in (a), and exponent b of the power spectrum in (b).
From music sequence I (Fig. 2(a)), for example, two different scaling regions can be clearly identied in the
graph. Region with short-time scale designates that the interval between musical notes is no more than a few
bars, while large time scale means musical notes are tens or even hundreds of bars away. The values of the

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

423

Table 2
Values of the Hurst exponent H and exponent b of the power spectrum obtained from various music-walk sequences (large time-interval
analysis)
Large time-interval
analysis

II

IIIa

IV(a)

IV(b)

V(a)

V(b)

VI(a)

VI(b)

VII(a)

VII(b)

H
b
2H 1
Deviation

0.0175
0.93
1.035
10.14%

0.0005
1.08
1.001
7.89%

0.22
1.43
1.44
0.69%

0.0050
1.01
1.01
0.00%

0.0075
0.92
1.015
9.36%

0.0009
1.09
1.0018
9.03%

0.0082
0.91
1.0164
10.47%

0.0013
1.17
1.0026
16.70%

0.0069
1.09
1.0138
7.52%

0.0043
0.91
1.0086
9.78%

0.0020
1.10
1.004
9.56%

a
Music sequence III has only one scaling region; hence the data listed here for this particular sequence are the same as that from the
short time-interval analysis listed in Table 1.

Fig. 2. (a) Root-mean-square deviation diagram and the associated Hurst exponent for the music-walk sequence Gavotte; (b) power
spectrum diagram and the associated exponent b for the music-walk sequence Gavotte.

Fig. 3. (a) Root-mean-square deviation diagram and the associated Hurst exponent for the music-walk sequence Le Cygne; (b) power
spectrum diagram and the associated exponent b for the music-walk sequence Le Cygne.

Hurst exponent for the graph in these two sections are 0.35 and 0.0175, respectively. The Hurst exponent is
not equal to 0.5, implying a certain correlation between increments of successive data values in the sequence.
For music sequences, these increments represent the difference in pitch between adjacent notes, which is
commonly referred to as the melody of the music. Thus, it is reasonable to assert that music piece I exhibits

ARTICLE IN PRESS
424

Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

Fig. 4. (a) Root-mean-square deviation diagram and the associated Hurst exponent for the music-walk sequence Ave Maria; (b) power
spectrum diagram and the associated exponent b for the music-walk sequence Ave Maria.

an anti-persistent correlation in melody in both short-scale and large-scale measures. On the other hand, the
value of the Hurst exponent in the short-scale range (0.35) is somewhat closer to 0.5, indicating that the music
sequence still preserves some degree of randomness in melody (changes of pitches) over an interval of a bar or
two. Several bars later, the Hurst exponent drops to a value near 0, meaning that the melody of the music
sequence exhibits a nearly perfect long-range correlation over the entire movement. That is, the melodic
motion of the music piece at one time is still correlated with that in the long past before. This might perhaps
reect the underlying structure of all musicbeing full of variation in short measures while still preserving
coherency in the long run. Our present analysis clearly discloses this genuine feature of music.
From Fig. 3(a) it can be observed that the H value (0.38) of music II in the short-time scale region is
relatively larger than that of music I. As the Hurst exponent H and the fractal dimension D of the curve are
related by equation D 2  H, a larger Hurst exponent implies a smoother music sequence, xn, and less
obvious uctuation in pitch over the examined time period. This can be proven by inspecting Fig. 1 where the
curve of music II appears smoother locally than that of the previous one. As in music I, there are two different
scaling regions in music II. The Hurst exponent corresponding to the large-time scale is 0.0005 (see the right
part of Fig. 3(a)), showing again a nearly perfect long-range correlation in the melodic motion. Note in this
diagram that the mild oscillations at the end of the curve indicate repetitions of measures in this musical piece,
suggesting a periodic feature in the structure of the music sequence. Admittedly, it has already been observed
from Fig. 1 that music II is more cyclic than music I.
Unlike previous two pieces of music, music sequence III shows only one scaling region (see Fig. 4(a)), and its
Hurst exponent value is about 0.22, the smallest among three. That music III has the smallest Hurst exponent
among the three works implies its sequence xn is more rugged in appearance, with more noticeable
uctuations in pitch. Furthermore, the Hurst exponent 0.22 is further away from the value 0.5, indicating that
music sequence III is somewhat less random in pitch changes (i.e., shows greater correlation in the increments
of adjacent values in the sequence), and hence lacks of diversity in melodic motion than the aforementioned
two music sequences I and II. In fact, among the three musical pieces, music III sounds relatively monotonous.
Therefore, the melody of a musical work can indeed be quantied and described by calculating the Hurst
exponent of its score.
The power spectra of these three musical pieces are given in part (b) of Figs. 24. Basically, all energy
spectra show the power-law 1=f b tendency with b40. The gradual increase of the spectral energy toward lowfrequency end of the spectrum is caused by the long-tailedness effect (slowly decaying autocorrelation
function) of the music sequencea typical behavior of a scaling noise with long-range correlation. Closer
inspection suggests that both music sequences I and II also display two different scaling regions in their
spectra, similar to that exhibited by the sDx curve. The dividing frequency of these two regions can be

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

425

estimated by locating the turning point of the sDx curve (the location Dnc where the sDx curve changes its
slope); then the dividing frequency is roughly 1=Dnc . The right side of the sDx diagram represents region
with large time scale and hence corresponds to the low frequency part of the spectrum diagram; likewise, the
left side of the sDx diagram is associated with the high frequency part of the spectrum diagram. Either way,
the b value calculated from both high and low frequency parts of the spectrum approximately satises the
equation b 2H 1 (see Tables 1 and 2), indicating that music sequence indeed possesses the properties
of an fBm.
The rest of the music pieces IV(a,b)VII(a,b) studied in this paper are all duos. Results are summarized in
Table 1 for short time-interval (high-frequency) analysis and Table 2 for large time-interval (low-frequency)
analysis. It can be ascertained from these tables that the corresponding H and b values of the two musical
parts in a duet are very close. The result suggests that as composers compile the two music parts in a duet,
eventually a similar fractal structure and dimension would be formed. The theoretical reason for this is not
yet clear. The results here may be coincidental and analyses of more samples are required for validation.
Additionally, from Tables 1 and 2 it can be observed that both the H value and the b value satisfy the
relation b 2H 1 well, which is demonstrated by the minor deviation percentages shown in the last rows
of the tables.
Noticeably, the Hurst exponent values of all music-walk sequences analyzed in this study are smaller than
0.5 (an anti-persistent trend in melodic motion), and this may be a typical characteristic of music in general. It
is doubtful whether a music score with a Hurst exponent far greater than 0.5 exists at all, because a Hurst
exponent over 0.5 indicates persistency in the melodic motion of the music; that is, the tones of adjacent
musical notes are well too positively correlated, making the music lacking in surprises and sound boring.
The fact that the spectrum exponent b of each music sequence ranges between 0 and 2 reveals that music is
neither unrelated random noise nor Brownian noise, with the strongest correlation, but is something else inbetween, which integrates randomness and orderliness in tones arrangement. This echoes the properties of an
anti-persistent fBm. In addition, it can be found that almost all musical works analyzed in this paper exhibit
two different scaling regions, implying that same piece of music would differ in correlation and characteristic
under different time scales. This is quite understandable. We have all possibly experienced a certain piece of
music in which several measures demonstrate considerable variation while the whole movement has a
consistent musical style.
Examining the b values for all music pieces studied here, it can be seen that the results are remarkably
consistent, 1:4obo1:8 for high-frequency area of the spectra. The values are somewhat closer to the
Brownian-noise value (b 2, which corroborates the results of Boon and Decroly [30]. This observation also
conrms the qualitative results of Nettheim based on the analysis of ve melodic lines (Bach, Mozart,
Beethoven, Schubert and Chopin) [31]. Nevertheless, in the low-frequency area, the b values eventually all turn
out to be close to 1, that of a typical 1=f -noise having perfect long-range correlation.
Finally, it has been mentioned before that two pitches are in harmony if they have a ratio of small integers.
Therefore, when two tones are being played simultaneously, it is the pitch difference between them that
determines whether or not they are in harmonyfor example, the combination of two tones with pitch
interval i 7 (a perfect fth) is sonorous, while a difference of pitch interval i 6 (a diminished fth) is
dissonant. Therefore, the sequences of pitch differences obtained by deduction between the two musical parts
in the works of Beethoven, Mozart, Pachelbel, and Bach (music sequences IV(a,b)VII(a,b)) are processed
again with the above Hurst exponent and power spectral analyses. The results from the Beethovens work
(music IV(a,b)) are shown in Fig. 5. Since the sequence derived from the pitch differences between two musical
parts also displays the feature of power-law scaling and satises the equation b 2H 1, it can be deduced
that the progression of harmonic in a duet is also of a fractal nature. Values of respective Hurst exponent H
and frequency spectrum exponent b are given in Tables 3 and 4.
5. Conclusion
The fractal nature of music has been constructed by applying the Hurst exponent and Fourier spectral
analyses to the music-walk sequences converted from various music scores. Both the root-mean-square
deviation and the power spectrum of the music-walk sequences exhibit a power-law scaling property, a

ARTICLE IN PRESS
426

Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

Fig. 5. (a) Root-mean-square deviation diagram and the associated Hurst exponent for the sequence deduced from the pitch differences
between two musical parts of Beethovens work Op. 24 No. 5; (b) power spectrum diagram and the associated exponent b for the sequence
deduced from the pitch differences between two musical parts of Beethovens work Op. 24 No. 5.

Table 3
Values of the Hurst exponent H and exponent b of the power spectrum for the sequence deduced from the pitch differences between two
musical parts of various duets (short time-interval analysis)
Short time-interval analysis

IV(a)IV(b)

V(a)V(b)

VI(a)VI(b)

VII(a)VII(b)

H
b
2H 1
Deviation

0.29
1.60
1.58
1.27%

0.23
1.44
1.46
1.37%

0.33
1.51
1.66
9.04%

0.14
1.26
1.28
1.56%

Table 4
Values of the Hurst exponent H and exponent b of the power spectrum for the sequence deduced from the pitch differences between two
musical parts of various duets (large time-interval analysis)
Large time-interval analysis

IV(a)IV(b)

V(a)V(b)

VI(a)VI(b)

VII(a)VII(b)

H
b
2H 1
Deviation

0.0034
0.99
1.0068
1.67%

0.0029
1.04
1.0058
3.40%

0.0102
1.15
1.0204
12.70%

0.0152
1.03
1.0304
0.04%

resemblance to the long-range correlation found in many naturally occurring uctuating phenomena. Music is
self-similar, both in audition (by listening to the successive changes in the acoustic frequency) and in visual
representation (by looking at the up-and-down uctuation of notes on the staffs). The melodic arrangement of
the entire movement of a music score is, to some extent, similar to the tone variation in just a few bars. The
lack of a coherent fractal structure in a random uctuation is one of the key points that essentially discern
noise from music.
Closer investigation shows that music-walk sequences generally display a scaling property in two apparently
different temporal scales with two different Hurst exponents. Short time-interval analysis gives typical Hurst
exponent values H 0:220:4; meaning that the tone change in music is normally anti-persistent in trend while
at the same time still preserving certain degree of randomness in its progression over a few measures. Large
time-interval analysis gives H  0, which results in a near 1=f -type power spectrum, signifying a perfect longrange correlation of melodic motion over the entire musical movement. This may reect the basic idea about

ARTICLE IN PRESS
Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

427

musictone variation in short measures makes the music abundant in expression, while long-term correlation
connects hundreds of bars of different melodious messages into a coherent piece of music. The fact the
exponents of the power spectra of all music-walk sequences studied here range between 0 and 2 once again
makes evident that music integrates both randomness and certainty like many other natural phenomena do.
It is also found that the Hurst exponent values of the musical works analyzed in this study are all smaller
than 0.5, indicating that music can be commonly treated as an anti-persistent fBm. This may be a typical
characteristic of music in general. Because a Hurst exponent greater than 0.5 implies the persistency of tone
changes in music, one part of music would be too similar to other parts in melodic motion, making the music
lacking in variability and sound monotonous.
The way in which music communicates with the mind may well not be realized in entirety until we achieve a
thorough understanding of how humans mentality, sense and sensitivity are commanded by the brain. We do,
however, begin to gain knowledge about the universality of the underlying structures embedded in diverse
works of nature. If music is invented to mimic a certain harmony in nature, then mountains are songs; rivers
are lyrics. Modern musicians use the concepts of fractals and chaos in their compositions [3237]; music
composed in this way must of course be part of the fractal scene.
References
[1] C. Madden, Fractals in Music: Introductory Mathematics for Musical Analysis, High Art Press, Salt Lake City, 1999.
[2] K.J. Hsu, Fractal geometry of music: from bird songs to Bach, in: A.J. Crilly, R.A. Earnshaw, H. Jones (Eds.), Applications of
Fractals and Chaos, Springer, New York, 1993, pp. 2139.
[3] B.B. Mandelbrot, The Fractal Geometry of Nature, Freeman, New York, 1983.
[4] R.F. Voss, J. Clarke, Nature 258 (1975) 317.
[5] R.F. Voss, J. Clarke, J. Acoust. Soc. Am. 63 (1) (1978) 258.
[6] B.J. West, M.F. Shlesinger, Am. Sci. 78 (1990) 40.
[7] R.F. Voss, Fractals in nature: from characterization to simulation, in: H.-O. Peitgen, D. Saupe (Eds.), The Science of Fractal Image,
Springer, New York, 1988, pp. 2170.
[8] M. Schroeder, Nature 325 (1987) 765.
[9] P. Campbell, Nature 325 (1987) 766.
[10] M. Schroeder, Fractals, Chaos, Power Laws: Minutes from an Innite Paradise, Freeman, New York, 1990.
[11] M. Gardner, Scientic American 238 (4) (1978) 16 (also see M. Gardner, Fractal Music, Hypercards, W.H. Freeman, New York,
1992).
[12] A.T. Scarpelli, Per. Comput. 3 (7) (1979) 17.
[13] D.E. Thomsen, Sci. News 117 (1980) 187.
[14] C. Dodge, C.R. Bahn, Byte June (1986) 185.
[15] C. Dodge, Comput. Music J. 12 (3) (1988) 10.
[16] K.J. Hsu, A.J. Hsu, Proc. Natl. Acad. Sci. USA 87 (1990) 938.
[17] K.J. Hsu, A.J. Hsu, Proc. Natl. Acad. Sci. USA 88 (1991) 3507.
[18] Y. Shi, Fractals 4 (1996) 547.
[19] M. Bigerelle, A. Iost, Chaos Soliton Fract. 11 (2000) 2179.
[20] G. Madison, J. New Music Res. 29 (2000) 335.
[21] B. Manaris, D. Vaughan, C. Wagner, J. Romero, R.B. Davis, Lecture Notes in Computer Science, vol. 2611, Springer, Berlin, 2003,
p. 522.
[22] G. Gunduz, U. Gunduz, Physica A 357 (2005) 565.
[23] B.B. Mandelbrot, J.W. Van Ness, SIAM Rev. 10 (4) (1968) 422.
[24] F. Reif, Fundamentals of Statistical and Thermal Physics, McGraw-Hill, New York, 1965.
[25] D.L. Turcotte, Fractals and Chaos in Geology and Geophysics, second ed., Cambridge University Press, New York, 1997.
[26] R.M. Crownover, Introduction to Fractals and Chaos, Jones and Bartlett, Boston, 1995.
[27] R.F. Voss, Physica D 38 (1989) 362.
[28] R.F. Voss, Random fractals: characterization and measurement, in: F. Family, T. Vicsek (Eds.), Dynamics of Fractal Surfaces, World
Scientic, Singapore, 1991, pp. 3950 (also see R.F. Voss, in: R. Pynn, A. Skjeltorp (Eds.), Scaling Phenomena in Disordered
Systems, W.H. Freeman, New York, 1985, pp. 111).
[29] R.F. Voss, Random Fractal Forgeries, in: J.E. Bresenham, R.A. Earnshaw, M.L.V. Pitteway (Eds.), Fundamental Algorithms for
Computer Graphics, Springer, New York, 1991, pp. 805835.
[30] J.P. Boon, O. Decroly, Chaos 5 (1995) 501.
[31] N. Nettheim, J. New Music Res. 21 (1992) 135.
[32] B. Fagarazzi, A fractal approach to musical composition, in: H.-O. Peitgen, J.M. Henriques, L.F. Penedo (Eds.), Fractals in the
Fundamental and Applied Sciences, Elsevier Science, New York, 1991, pp. 135146.

ARTICLE IN PRESS
428

Z.-Y. Su, T. Wu / Physica A 380 (2007) 418428

[33]
[34]
[35]
[36]
[37]

P. Campbell, Nature 324 (1986) 523.


M. Schroeder, Fractals in music, in: C.A. Pickover (Ed.), Fractal Horizons, St. Martins Press, New York, 1996, pp. 207223.
J. Pressing, Comput. Music J. 12 (2) (1988) 35.
M. Gogins, Comput. Music J. 15 (1) (1991) 40.
R. Bidlack, Comput. Music J. 16 (3) (1992) 33.